A stereo calibration method and device adapted to high-speed binocular vision scenarios

Abstract

This invention relates to a stereo calibration device adapted to high-speed binocular vision scenarios, comprising an L-shaped calibration plate body, which is composed of two perpendicularly intersecting calibration plates, and a preset calibration pattern composed of multiple self-illuminating calibration points is provided on the calibration surfaces of the two calibration plates; a control module, electrically connected to the self-illuminating calibration points, is used to control the emission timing of the self-illuminating calibration points, so that their flashing frequency is synchronized with the shooting frame rate of the camera in the high-speed binocular vision system. This invention also provides a stereo calibration method adapted to high-speed binocular vision scenarios based on the above device. This invention adapts to the short exposure and high frame rate requirements of high-speed cameras, eliminates the need for supplementary lighting and moving the calibration plate, and filters high-quality calibration images through intra-modal and inter-modal dual consistency indices, effectively improving the accuracy of calibration parameters while completing calibration quickly.

Description

Technical Field

[0001] This invention belongs to the field of calibration technology for high-speed binocular vision measurement systems, and specifically relates to a stereo calibration method and device adapted to high-speed binocular vision scenarios. Background Technology

[0002] Binocular vision measurement is an important non-contact measurement method, but it faces many challenges when dealing with high-speed target scenes. High-speed cameras typically have high frame rates, exceeding 1000fps, requiring short exposure times (below 1ms) to avoid motion blur. This results in low brightness and poor signal-to-noise ratio in images captured under normal lighting conditions. Furthermore, during high-speed continuous acquisition, underexposure, lighting fluctuations, or sensor noise can easily lead to quality issues such as calibration point spot diffusion, adhesion, or abnormal shapes, significantly reducing the detection stability of calibration points. In such cases, using conventional calibration boards often requires reducing the frame rate, increasing exposure, and adding extra light, making the process cumbersome and time-consuming. Moreover, existing calibration methods often rely on planar calibration boards such as checkerboards or dot arrays, which not only require manual movement but also cannot adapt to complex and changing acquisition environments. They also struggle to establish a coordinate system based on the actual position of the calibration board, making them unsuitable for high-speed binocular vision scenarios. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies by providing a stereo calibration method and apparatus adapted to high-speed binocular vision scenarios. It can be used with high-speed cameras operating at the same frequency, completing calibration in a very short time. Furthermore, the calibration process requires no adjustment of the calibration plate position. It can automatically evaluate image frame quality and select reliable calibration frames under high-speed vision conditions, significantly increasing calibration efficiency and accuracy. The specific technical solution is as follows: In a first aspect, the present invention provides a stereo calibration device adapted to high-speed binocular vision scenarios, the device comprising: The L-shaped calibration plate body is composed of two calibration plates that intersect perpendicularly, and the calibration surfaces of the two calibration plates are provided with a preset calibration pattern composed of multiple self-luminous calibration points; among the multiple self-luminous calibration points, there is a standard calibration point as a brightness reference, and there is a known and different difference between the luminous power of the other self-luminous calibration points and the luminous power of the standard calibration point. The control module is electrically connected to the self-illuminating calibration point and is used to control the emission timing of the self-illuminating calibration point so that its flashing frequency is synchronized with the shooting frame rate of the high-speed binocular camera.

[0004] As a preferred embodiment of the present invention, the self-luminous calibration point is an LED bead or a light-emitting element made of a light-emitting material.

[0005] As a preferred embodiment of the present invention, the plurality of self-illuminating calibration points include at least three standard calibration points for constructing a spatial three-dimensional coordinate system in the calibration image; including the center point located at the connection of two calibration plates, the edge point of the connection, and the center point of the edge of a calibration plate away from the connection.

[0006] As a preferred embodiment of the present invention, the control module dynamically designates a selected self-illuminating calibration point as a standard calibration point for constructing the spatial three-dimensional coordinate system by setting a luminous intensity higher than a predetermined brightness threshold.

[0007] As a preferred embodiment of the present invention, the preset pattern is a rectangular dot matrix arranged at equal intervals.

[0008] Secondly, the present invention also provides a stereo calibration method for high-speed binocular vision scenes based on the above-mentioned device, comprising the following steps: Step S1: Synchronously start the high-speed binocular camera and the stereo calibration device. The high-speed binocular camera acquires multiple frames of calibration images by combining different self-illuminating calibration points that are controlled by the control module to light up in a time-division manner. The control module controls the flashing frequency of the self-illuminating calibration points to be synchronized with the frame rate of the high-speed binocular camera. Step S2: Determine the position of the standard calibration point in the calibration image based on the gray value of the light spot, and determine the candidate region of each light emission calibration point based on the position of the standard calibration point and the arrangement structure of the self-emitting calibration points; Step S3: Evaluate the noise intensity of the current frame calibration image based on the distinguishability of the candidate region of each self-illuminating calibration point in the current frame calibration image and the observational consistency of the luminous power of all self-illuminating calibration points in the current frame calibration image; Step S4: Based on the noise intensity, select a reliable calibration subset from the multi-frame calibration images, and solve the calibration parameters of the high-speed binocular vision system based on the reliable calibration subset.

[0009] As a preferred embodiment of the present invention, step S3, which assesses the noise intensity of the current frame calibration image based on the distinguishability of the candidate region of each self-illuminating calibration point in the current frame calibration image and the observational consistency of the luminous power of all self-illuminating calibration points in the current frame calibration image, includes: Step S31: Calculate the intramodal consistency index of the current frame calibration image based on the grayscale information and spot morphology information of the self-illuminating calibration points and their neighboring self-illuminating calibration points in the current frame calibration image. The intramodal consistency index is used to characterize the degree of distinguishability between the self-illuminating calibration points and their neighboring self-illuminating calibration points in the current frame calibration image. Step S32: Construct an observed grayscale vector based on the grayscale values ​​of all self-illuminating calibration points in the current frame calibration image, and construct a prior power vector based on the difference in luminous power of all self-illuminating calibration points; calculate the cosine similarity between the observed grayscale vector and the prior power vector to obtain an intermodal consistency index, which is used to characterize the degree of observational consistency between the grayscale distribution of the current frame image and the luminous power distribution of the calibration board; Step S33: Construct noise intensity by weighted fusion of the intramodal consistency index and the intermodal consistency index.

[0010] As a preferred embodiment of the present invention, step S31, which involves calculating the intramodal consistency index of the current frame calibration image based on the grayscale information and spot morphology information of the self-illuminating calibration points and their neighboring self-illuminating calibration points in the current frame calibration image, includes: Step S311: Select a local background region outside the candidate region of each light emission calibration point, calculate the background mean based on the pixel gray value in the local background region, and estimate the local background noise variance based on the fluctuation of each pixel in the local background region relative to the background mean. Step S312: Normalize the peak gray value of the candidate region of the self-illuminating calibration point after background removal based on the background mean by using the local background noise variance to obtain the single-point signal-to-noise ratio of the self-illuminating calibration point. Step S313: Based on the minimum difference between the peak gray level of the candidate region of the self-illuminating calibration point and the peak gray level of the candidate region of its neighboring self-illuminating calibration points, the normalized value under the local background noise variance is obtained to obtain the neighboring point separation degree of the self-illuminating calibration point. Step S314: Collect all reference frames when the self-propelled cursor is fixed and lit, and obtain the spot shape anomaly degree based on the difference between the spot width of the self-propelled cursor fixed point in the current frame and the average spot width of the reference frame. Step S315: Aggregate the single-point signal-to-noise ratio, the adjacent point separation degree, and the spot shape anomaly degree to obtain the intramodal consistency index of the current frame calibration image.

[0011] As a preferred embodiment of the present invention, step S32, which involves constructing an observed grayscale vector based on the grayscale values ​​of all self-illuminating calibration points in the current frame calibration image, and constructing a priori power vector based on the differences in luminous power among all self-illuminating calibration points, includes: Step S321: In the current frame calibration image, remove the background from the peak gray value of each self-propelled cursor point and normalize it through the standard calibration point. Then, concatenate the peak gray values ​​of all the self-propelled cursor points after normalization as the observed gray value vector. Step S322: Normalize the luminous power of each self-firing cursor point to the luminous power of the standard calibration point, and concatenate the luminous power of all the normalized self-firing cursor points in sequence as a priori power vector.

[0012] As a preferred embodiment of the present invention, step S34, which involves filtering and obtaining a reliable calibration subset from multiple calibration images based on the noise intensity, includes: The noise intensity of the calibration image is statistically modeled frame by frame using a Gaussian mixture model, which includes at least two Gaussian distribution components. One component represents the noise intensity distribution of reliable frames, and the other component represents the noise intensity distribution of noisy frames. The posterior probability of each calibration image belonging to a reliable calibration subset is obtained by calculating the probability density of the noise intensity of each frame under each component.

[0013] The beneficial effects of this invention are: 1. The calibration board of the present invention can work at the same frequency as a high-speed camera and complete the calibration image acquisition in a very short time. At the same time, the calibration board can emit light on its own and is suitable for high-speed short exposure experimental scenarios. 2. The calibration plate of the present invention has a non-planar structure. During the calibration process, there is no need to change the position of the calibration plate. The mapping relationship between the spatial coordinates of the calibration points and the image coordinates can be obtained based on the positional relationship of different calibration points to complete the calibration. 3. At least three standard calibration points in the calibration plate of the present invention can be used to automatically identify and generate a coordinate system. After calibration is completed by this method, the resulting spatial coordinate system is located at the position of the calibration plate, which facilitates subsequent data processing.

[0014] 4. By constructing an intramodal consistency index, we can effectively identify situations such as spot adhesion or abnormal morphology, thereby improving the reliability of calibration point localization. Based on the intermodal consistency index between the observed brightness vector and the prior power vector, we can detect calibration point mismatches or grayscale structure anomalies, improving the accuracy of calibration point matching. Finally, based on frame-by-frame noise identification and using a Gaussian mixture model to statistically model the noise loss distribution, we can achieve adaptive probability determination of reliable and noisy frames, avoiding misjudgment problems caused by fixed thresholds. This allows high-quality images to play a greater role in calibration calculations, thereby improving the stability and accuracy of binocular camera calibration results. Attached Figure Description

[0015] Figure 1 A schematic diagram of the calibration device of the present invention is shown; Figure 2 Two comparative schematic diagrams are shown in the calibration device of the present invention, illustrating the dynamic designation of the standard calibration point position. Figure 3The diagram illustrates eight emission modes of calibration points during the multi-frame calibration image acquisition process in the calibration method of this invention. Figure 4 A flowchart of the calibration method of the present invention is shown. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] Example 1 To address the technical problems in the background section, a stereo calibration device adapted for high-speed binocular vision scenarios is provided as follows: Combination Figure 1 As shown, a stereo calibration device adapted to high-speed binocular vision scenarios includes an L-shaped calibration plate body, which is composed of two calibration plates that intersect perpendicularly, and the calibration surfaces of the two calibration plates are provided with a preset calibration pattern composed of multiple self-illuminating calibration points.

[0018] The control module (not shown in the figure) is electrically connected to the self-illuminating calibration point and is used to control the emission timing of the self-illuminating calibration point so that its flashing frequency is synchronized with the shooting frame rate of the camera in the high-speed binocular vision system.

[0019] In the above technical solution, the L-shaped calibration plate body provides two mutually perpendicular calibration planes at once, forming a three-dimensional spatial reference that can be directly used for the calibration of binocular systems. This avoids the tedious operation of repeatedly moving and placing the planar calibration plate and simplifies the calibration process. The control module synchronizes the blinking of the calibration point with the camera frame rate, ensuring that a clear calibration point image can still be captured in a short exposure time of milliseconds, thus adapting to high-speed shooting scenarios.

[0020] The aforementioned multiple self-illuminating calibration points are arranged to form a preset calibration pattern, which is used by a high-speed binocular camera to capture calibration features in order to complete the calibration of the high-speed binocular vision measurement system.

[0021] The aforementioned L-shaped calibration plate body can be composed of two calibration plates vertically fixedly connected, or it can be composed of two calibration plates rotatably connected, allowing the calibration plate to switch between an unfolded state and a folded state. In the unfolded state, the two calibration plates maintain a 90° perpendicular intersection to construct a three-dimensional calibration reference surface. In the folded state, the two calibration plates fit together for easy storage and carrying. The aforementioned rotatable connection structure is existing technology and includes, but is not limited to, damping shaft structure, hinge structure, slide groove and shaft combination structure, or magnetic adsorption connection structure. Specific structures will not be described in detail here.

[0022] The dimensions of the two calibration plates are adapted to the overlapping area of ​​the dual-camera field of view of the high-speed binocular vision measurement system, so that full field of view calibration can be completed without moving the calibration plates.

[0023] Preferably, the control module is typically a synchronous controller, an existing precise timing control device. It can simultaneously emit two signals: one to control the high-speed camera to capture images, and the other to control the calibration board to emit light. The time difference between these two signals is generally less than 1 ns. Because the high-speed camera has an extremely high frame rate, typically above 1000 FPS, and the calibration board's emission time is also on the microsecond level, without this device, it would be very difficult to manually control the moment the calibration board emits light so that the high-speed camera can simultaneously capture an image. Therefore, a control module is needed to simultaneously send two signals, ensuring that the calibration board and the high-speed camera operate at the same time.

[0024] Preferably, the self-illuminating calibration point is an LED bead or a light-emitting element made of a light-emitting material.

[0025] In the above technical solution, active light-emitting elements are used, which enables the calibration board to generate feature points with sufficient brightness without relying on external supplementary lighting equipment under low ambient light or high-speed camera short exposure conditions. This enhances the environmental adaptability and reliability of the calibration process and ensures the quality of image acquisition.

[0026] Among the multiple self-luminous calibration points, there is a standard calibration point that serves as a brightness reference, and there is a known and distinct difference between the luminous power of the remaining self-luminous calibration points and the luminous power of the standard calibration point.

[0027] In the above technical solution, by presetting a unique and known luminous power for each calibration point, different luminous intensities are generated, so each point presents a unique grayscale value in the image. This brightness encoding scheme enables the system to automatically and uniquely identify each calibration point in the image simply by analyzing the grayscale value, without relying on the geometry of the dot matrix or complex visual patterns, thus achieving efficient and accurate matching of the calibration point image coordinates and physical space coordinates.

[0028] like Figure 1 As shown, preferably, the standard calibration point uses a green indicator LED, which has the standard power. The power of the other LEDs varies, but their power relative to the standard LED is known. During calibration, the calibration board power supply is turned on (this power supply powers the calibration point; this type of power supply is existing technology and will not be described in detail here). The LEDs emit light, and a high-speed camera captures images of the calibration board. Since the power of all LEDs on the calibration board, except for the standard power LEDs, is different, the grayscale value of each LED in the captured calibration board image is different. Based on the grayscale value of each LED, the corresponding position of each LED on the calibration board can be identified. The image coordinates of the LEDs can then be correlated with their actual spatial coordinates to obtain the calibration parameters and complete the calibration.

[0029] The L-shaped calibration board can work synchronously with the high-speed camera used for image acquisition. The LEDs on the calibration board blink according to the set parameters, and their blinking frequency is the same as the shooting frequency of the high-speed camera. For example, if the sampling frequency of the high-speed camera is 1000fps, then the blinking frequency of the LEDs on the calibration board is set to 1000fps. If 30 frames of calibration images are acquired, the entire process of acquiring calibration images can be completed within 1 second.

[0030] Among the multiple self-illuminating calibration points, at least three are standard calibration points, used to construct a spatial three-dimensional coordinate system in the calibration image.

[0031] In the above technical solution, at least three standard calibration points are further set on the basis of brightness encoding. These points not only participate in coordinate matching, but also serve as the reference for constructing the world coordinate system. This allows a clear three-dimensional coordinate system to be formed directly on the physical space established by the calibration plate after calibration, facilitating subsequent direct and intuitive three-dimensional position and attitude measurement of the measured object, thus improving the practicality of the measurement system.

[0032] Preferably, such as Figure 1 As shown, there are three standard calibration points, including the center point at the junction of the two calibration plates, the edge point of the junction, and the center point of the edge of one calibration plate away from the junction. Preferably, the three standard calibration points are three standard power LED beads.

[0033] The above technical solution provides a preferred and fixed standard calibration point layout scheme. Selecting locations with clear geometric features, such as the center and edges of the connection points, provides a stable, repeatable, and easily understood physical reference for coordinate system construction. This layout closely links the coordinate system to the physical structure of the calibration plate, ensuring clear positioning, reducing ambiguity in coordinate system definition, and facilitating alignment and comparison between different calibration results.

[0034] Preferably, the preset pattern is a rectangular dot matrix with equidistant and uniform arrangement; the rectangular dot matrix is ​​preferably a 7-row × 7-column dot matrix.

[0035] In the above technical solution, the preset pattern is limited to a rectangular dot matrix arranged at equal intervals. This regular and symmetrical pattern design facilitates the processing and manufacturing of the calibration plate, reducing production costs. At the same time, the regular arrangement gives the spatial coordinates of the calibration points a simple mathematical relationship (such as row and column numbers, which can be used to deduce the positions of other calibration points based on the standard calibration points), simplifying the establishment and management of the preset spatial coordinate database and improving the feasibility of the system.

[0036] Example 2 Combination Figure 2As shown, based on the above embodiments, this embodiment further provides the following: In this embodiment, the control module dynamically designates a selected self-illuminating calibration point as a standard calibration point for constructing the spatial three-dimensional coordinate system by setting a luminous intensity higher than a predetermined brightness threshold.

[0037] The above technical solution breaks through the limitation of fixed reference point positions in traditional calibration plates. By dynamically setting any selected calibration point as a highlighted standard point through a control module, users can flexibly and customarily select the physical positions of the coordinate system origin and axes according to specific measurement task requirements. This flexibility allows the measurement coordinate system to be better aligned with the measured object or area of ​​interest, reducing unnecessary coordinate transformations and improving the convenience and intuitiveness of the measurement.

[0038] Specifically, in this solution, the positions of the three standard calibration points can be selected according to the needs of the actual application scenario. If the customer needs to construct coordinate systems at different locations, they can choose different calibration points as the three standard points. Traditionally, the establishment of a calibration board coordinate system is achieved by identifying outliers on the calibration board. Once the design of the calibration board is determined, the positions of these outliers cannot be changed. Therefore, it is impossible to specify the location of the coordinate system origin according to customer needs. In this solution, however, the requirement for the three standard calibration points can be selected according to customer requirements, and the identification of the three standard calibration points is determined based on brightness. Therefore, we can control the brightness of different calibration points on the calibration board through a control module, setting a brightness threshold for the standard points. When the brightness of a calibration point on the calibration board reaches the set threshold, it is considered a standard calibration point, and the coordinate system is set based on the standard calibration point. This solution allows for the construction of different coordinate systems by changing the positions of the standard calibration points according to customer requirements, facilitating the customer's acquisition of the relative positional relationship of the tested object within the coordinate system.

[0039] Example 3 Combination Figures 1-4 As shown, based on the above embodiments, this embodiment further provides the following: In this embodiment, as Figure 2 and Figure 4 As shown, a stereo calibration method adapted to high-speed binocular vision scenes based on the above-mentioned device includes the following steps: Step S1: Synchronously start the high-speed binocular camera and the stereo calibration device. The high-speed binocular camera acquires multiple frames of calibration images by combining different self-illuminating calibration points that are controlled by the control module to light up in a time-division manner. The control module controls the flashing frequency of the self-illuminating calibration points to be synchronized with the frame rate of the high-speed binocular camera.

[0040] like Figure 3 and Figure 4As shown, multiple frames of calibration images are acquired by illuminating different combinations of self-illuminating calibration points at different times to form an initial calibration set. Illuminating different subsets of calibration points at different times effectively reduces mutual interference between light points in a single frame image, improving the clarity and independence of each feature point imaging. Acquiring multiple frames of calibration images allows for the fusion of calibration point detection and matching results from multiple frames, utilizing more observation data to adjust and optimize calibration parameters, thereby effectively improving the overall accuracy of the final calibration result.

[0041] Specifically, the calibration image acquisition process is as follows: Figure 3 The diagram illustrates various methods for acquiring emission data at calibration points. White points represent emission points, and all three standard calibration points emit light in each acquisition. This method allows for the acquisition of calibration information from different locations within the calibration board's spatial location. By comparing multiple sets of calibration results, a higher precision calibration outcome can be obtained. It should be noted that... Figure 3 These are merely preferred exemplary solutions. In actual implementation, adjustments can be made as needed, and the solutions are not limited to the eight lighting methods shown in the figure.

[0042] Step S2: Determine the position of the standard calibration point in the calibration image based on the gray value of the light spot, and determine the candidate region of each luminous calibration point based on the position of the standard calibration point and the arrangement structure of the self-luminous calibration points.

[0043] In each frame of the image, this step determines the reference coordinate system of the calibration board in the current image by detecting the position of the standard calibration points, and establishes candidate regions for each luminous calibration point according to the preset spatial topology of the calibration board (i.e., the arrangement interval and order between each self-illuminating calibration board on the calibration board). Specifically, The standard calibration points utilize the highest luminous power. The pixel positions of the three standard calibration points in the image are determined by retrieving the maximum grayscale value in the calibration image. A reference coordinate system is constructed based on the three standard calibration points. The approximate positions of the remaining calibration points in the image are predicted based on the pre-set dot matrix structure of the calibration board, and candidate regions are set around the predicted positions. These steps determine the search area for each calibration point, thereby improving subsequent detection efficiency.

[0044] Assume the calibration board has M light-emitting points, and the three-dimensional coordinates of each point in the calibration board coordinate system are: .

[0045] In each frame of the calibration image, a candidate region is generated for each calibration point. , indicating the first The set of candidate pixel regions for point n in the image. The nth point... Frame image denoted as , To calibrate the pixel coordinates on the image.

[0046] Step S3: Evaluate the noise intensity of the current frame calibration image based on the distinguishability of the candidate region of each self-illuminating calibration point in the current frame calibration image and the observational consistency of the luminous power of all self-illuminating calibration points in the current frame calibration image.

[0047] As a preferred embodiment, step S3 includes the following sub-steps: Step S31: Calculate the intramodal consistency index of the current frame calibration image based on the grayscale information and spot morphology information of the self-illuminating calibration points and their neighboring self-illuminating calibration points in the current frame calibration image. The intramodal consistency index is used to characterize the degree of distinguishability between the self-illuminating calibration points and their neighboring self-illuminating calibration points in the current frame calibration image.

[0048] It should be noted that grayscale information and spot morphology information of the luminous calibration points are extracted in each candidate region, including: peak grayscale, spot width, and spot area. Since noise in high-speed binocular systems is usually not uniform across the entire image—for example, different vignetting angles at the lens edges, varying background distribution; crosstalk or scattering near certain light points; localized reflections at certain locations; and inconsistent responses in different areas of the left and right cameras—each point and each frame has its own local background noise statistics. Specifically, the background information around each self-illuminating calibration point is calculated through the following steps: Step S311: Select a local background region outside the candidate region of each light emission calibration point, calculate the background mean based on the pixel gray value in the local background region, and estimate the local background noise variance based on the fluctuation of each pixel in the local background region relative to the background mean.

[0049] In one implementation, a reference frame can be acquired when all self-propelled cursor points are illuminated. The average area of ​​the light spot at each calibration point is calculated within the reference frame and used as the reference region for the candidate region. It should be noted that this operation is also performed on-site during calibration to avoid the influence of the actual measurement environment on the light spot size.

[0050] In this embodiment, the radius of the recording reference area is With the center of the candidate region of the current calibration point as the center, the radius is... The area within the radius is used as the main region of the light spot and is not used to estimate the background; arrive ( The area between the radius of the current candidate region and the background ring (i.e., the local background region) is defined as the background ring.

[0051] The pixels of this local background region are recorded as follows: , , .in, Indicates the first Frame in pixel grayscale value at that location Indicates a local background area. This represents the number of pixels in a local background region. The average background value... for: Then calculate the background noise variance : in, Indicates the first Frame, point Neighborhood background noise variance. During high-speed, short-exposure exposures, even without a real signal, background grayscale will fluctuate. This fluctuation originates from readout noise, dark current, ambient stray light, photon statistical fluctuations, etc. By statistically analyzing the local background noise level, we can more accurately assess the significance of the current calibration point signal intensity relative to the background noise, thus measuring the credibility of the candidate region itself.

[0052] Step S312: Normalize the peak grayscale of the candidate region of the self-illuminating calibration point after background removal based on the background mean by using the local background noise variance to obtain the single-point signal-to-noise ratio of the self-illuminating calibration point.

[0053] This step involves constructing the single-point signal-to-noise ratio (SNR) metric. For example, the single-point SNR can be calculated using the following formula: in, The peak gray level within the candidate region. This represents the standard deviation of the background noise.

[0054] Understandably, when the point signal is weak and the noise is strong under short exposure, the signal-to-noise ratio of a single point can directly reflect the reliability of the calibration point positioning.

[0055] Step S313: Based on the minimum difference between the peak gray level of the candidate region of the self-illuminating calibration point and the peak gray level of the candidate region of its neighboring self-illuminating calibration points, the normalized value under the local background noise variance is obtained to obtain the neighboring point separation degree of the self-illuminating calibration point.

[0056] In this embodiment, the eight-neighborhood of each self-illuminating calibration point is taken as its neighboring self-illuminating calibration points. For example, the separation degree between neighboring points can be calculated using the following formula: in, Indicates calibration point The nearest neighbor set is taken in this embodiment as the eight-neighbor set. This represents the difference in peak values ​​between adjacent points; The synthesized noise scale indicates that the noise level is more difficult to distinguish.

[0057] In high-speed, short-exposure scenarios, multiple light spots can interfere with each other, causing phenomena such as diffusion, ghosting, and sticking. When multiple light spots stick together, resulting in similar gray levels, This will significantly reduce the number of frames, thus enabling the identification of indistinguishable frames.

[0058] On the other hand, high-speed jitter, motion blur, and abnormal exposure can change the shape of the light spot, causing the center positioning to shift or multiple peaks to appear in the candidate area.

[0059] Step S314: Obtain the spot shape anomaly degree based on the difference between the spot width of the self-firing cursor in the current frame and the average spot width of the reference frame.

[0060] It should be noted that the reference frame is the calibration image when all self-illuminating calibration points on the calibration board are illuminating.

[0061] As an example only, this embodiment calculates the shape anomaly degree by the spot width. In other embodiments, the spot shape deviation can also be calculated by calculating the spot center offset, the spot aspect ratio, etc.

[0062] For example, this embodiment uses the following formula to calculate the spot shape anomaly degree: in, This represents the average spot width in the reference frame.

[0063] It should be noted that the above-mentioned calculation of the single-point signal-to-noise ratio, the adjacent point separation degree, and the spot shape anomaly degree does not restrict the order of the steps, that is, the above steps S312-S314 do not constitute a strict operation order.

[0064] Step S315: Aggregate the single-point signal-to-noise ratio, the adjacent point separation degree, and the spot shape anomaly degree to obtain the intramodal consistency index of the current frame calibration image.

[0065] For example, a monotonic mapping function can be used to compress the magnitude of the above three indicators to [0,1] to unify different dimensions to a comparable scale. Then, the average of the product of the above three indicators of all self-propelled cursor points in a frame of calibration image is aggregated as the intramodal consistency index of the current frame of calibration image. This can detect problems such as spot adhesion, grayscale diffusion, and difficulty in identifying calibration points.

[0066] It is understandable that intramodal consistency measures whether points within the same frame conform to the structural characteristics that should be distinguishable. This is a consistency index within a single modality of a single-frame calibration image.

[0067] Intramodal structural information cannot determine noise alone; it must also be considered in conjunction with intermodal alignment. Under the same exposure, the same frame, and the same imaging chain, the observed grayscale is usually the result of the luminous power after imaging mapping. If there is no severe adhesion, occlusion, or misidentification of the lamp points, then the relative magnitude and proportion of the grayscale at each point should maintain a consistent trend with the preset power distribution. However, in actual high-speed, short-exposure scenarios, the total brightness fluctuates greatly, resulting in inconsistent grayscale structures at various locations on the entire board. Therefore, in this embodiment, the known power fingerprint inherent in the calibration board is mapped onto the observed image to detect whether the trends of change between the two are consistent.

[0068] Step S32: Construct an observed grayscale vector based on the grayscale values ​​of all self-illuminating calibration points in the current frame calibration image, and construct a priori power vector based on the luminous power differences of all self-illuminating calibration points; calculate the cosine similarity between the observed grayscale vector and the priori power vector to obtain an intermodal consistency index, which is used to characterize the degree of observational consistency between the grayscale distribution of the current frame image and the luminous power distribution of the calibration board. Specifically, this step may include the following sub-steps: Step S321: In the current frame calibration image, remove the background from the peak gray value of each self-propelled cursor point and normalize it through the standard calibration point. Then, concatenate the peak gray values ​​of all the normalized self-propelled cursor points in order to form the observed gray value vector.

[0069] Under ideal or near-ideal conditions, the first The self-illuminating calibration point at the first The observed grayscale of a frame can be represented as: In the formula, For the first The self-illuminating calibration point at the first The observed grayscale of the frame; For the first The prior luminous power of a self-luminous calibration point; For the first The frame's imaging gain coefficient reflects the overall scaling effect, including exposure, lens transmittance, and sensor response. For the first The background offset of a frame can be reflected by the background grayscale of the calibrated image of that frame; For the first Frame noise includes Gaussian noise, photon noise, local perturbations, etc.

[0070] Therefore, the observed grayscale is not equal to the power, but rather the power after being mapped by the imaging system. Thus, background processing is required first, which in this embodiment is specifically performed using the following formula: In the formula, For the normalized first The observed grayscale of a self-illuminating calibration point For the first The self-illuminating calibration point at the first Peak gray levels in frame-calibrated images To calibrate the background grayscale of the image in frame t, The peak gray level of the standard calibration point in the t-th frame calibration image is given.

[0071] Then, by concatenating all the self-propelled cursor points in a single frame of the calibration image in their positional order, the observed grayscale vector is obtained. Step S322: Normalize the luminous power of each self-firing cursor point to the luminous power of the standard calibration point, and concatenate the luminous power of all the normalized self-firing cursor points in sequence as a priori power vector.

[0072] Similar to step S321 above, the prior luminous power is normalized: In the formula, For the first The normalized prior luminous power of a self-luminous calibration point is the ratio of the prior power to the standard point. The prior luminous power of the standard calibration point is a quantity that has been set and is known in advance on the calibration board.

[0073] Then, the prior power vector is obtained by concatenating all the self-propelled cursor points in a calibration image in positional order: The raw data from two modalities cannot be directly compared. Instead, they are first projected into the same comparable representation space, that is, each observation is first placed into a "candidate location slot" before forming a vector. Only in this way can the first and second vectors be compared. Each component represents the same light point at the same physical location, and then cosine similarity is used to measure directional consistency.

[0074] Then the intermodal consistency index for: in, It represents the inner product of the observed grayscale vector and the prior power vector, reflecting whether the corresponding components of the two vectors have the same trend of change. Indicates the length of the observed grayscale vector. This represents the length of the prior power vector.

[0075] It should be understood that intermodal consistency reflects the consistency index between the prior luminous power and the actual observed grayscale of two modes. Simply put, it can be understood as determining which points should be brighter and whether they are indeed brighter during observation. Whether the relative structure of the overall brightness distribution of a frame is consistent with the power distribution structure can accurately reflect the noise level, including factors such as camera hardware and ambient light.

[0076] Step S33: Construct noise intensity by weighted fusion of the intramodal consistency index and the intermodal consistency index.

[0077] In this embodiment, the noise intensity is constructed using the following formula: in, As an intramodal consistency index, As an intermodal consistency index, This is the balance coefficient.

[0078] Step S4: Based on the noise intensity, select a reliable calibration subset from the multi-frame calibration images, and solve the calibration parameters of the high-speed binocular vision system based on the reliable calibration subset.

[0079] Specifically, the noise intensity of the calibration image is statistically modeled frame by frame using a Gaussian mixture model. The Gaussian mixture model includes at least two Gaussian distribution components, one of which represents the noise intensity distribution of reliable frames and the other of which represents the noise intensity distribution of noisy frames. The posterior probability of each calibration image belonging to a reliable calibration subset is obtained by calculating the probability density of the noise intensity of each frame of calibration image under each component.

[0080] After obtaining the noise intensity of all frames, the image quality of each frame is evaluated based on its noise intensity. However, since the noise intensity of each frame drifts overall with exposure conditions, background noise, imaging gain, and local adhesion, the threshold discrimination method in conventional techniques does not have a stable and uniform demarcation meaning. Reliable frames and noisy frames usually overlap in the noise intensity domain, especially under short exposure and low signal-to-noise ratio conditions, making it difficult to accurately distinguish a large number of boundary frames using a fixed threshold.

[0081] Gaussian mixture models do not assume that noise loss in all frames comes from the same statistical mechanism, but rather that it comes from a mixture of several different statistical mechanisms. In this embodiment, a bimodal Gaussian mixture model is used, which can be expressed as: In the formula, Indicates noise intensity The overall probability density; This indicates the number of mixed components (it should be noted that, as mentioned above) (The meanings are different). In this embodiment, we take 2, which means that one component corresponds to the noise intensity distribution of a relatively reliable frame, and the other component is used to represent the noise intensity distribution of a noisy frame. Indicates the first Mixed weights of the components. Indicates the first The mean of each component, Indicates the first The variance of each component. After fitting with a bimodal Gaussian mixture model (which is existing technology and well known to those skilled in the art, and will not be elaborated further), the posterior probability that a certain frame belongs to a reliable component can be calculated.

[0082] A Gaussian mixture model is used to model the noise intensity distribution, which can adaptively estimate the reliable frame component and the noise frame component and output the frame-by-frame posterior probability, thereby improving the robustness of the reliable calibration subset selection and the accuracy of parameter calibration.

[0083] Then, parameter calibration begins. By using high-reliability calibration images from the selected reliable calibration subset, the pixel coordinates of the self-illuminating calibration points in the image are obtained, and their real-world three-dimensional spatial coordinates are acquired. Multiple points (at least six points) are used to form a mapping relationship, and the rotation matrix and translation vector of the high-speed binocular system are solved.

[0084] like Figure 1 and Figure 4 As shown, after the above steps, the system further includes constructing the three-dimensional spatial coordinate system based on the three standard calibration points, with one of the points serving as the origin.

[0085] The above technical solution clearly defines the specific construction method of the three-dimensional coordinate system. By selecting three standard calibration points and designating one of them as the origin, and utilizing the clear geometric principle that three points determine a plane and one of them is the origin, a right-handed rectangular coordinate system can be stably and uniquely determined. The coordinate coefficients constructed by this method have rigorous geometric definitions and clear physical meanings, ensuring that all subsequent three-dimensional measurement data have a consistent and reliable geometric interpretation, facilitating engineering applications.

[0086] A coordinate system is constructed based on the positions of three standard calibration points in each frame of calibration image. Specifically, the three standard calibration points are illuminated in each calibration image, and each frame of calibration image contains such a marker point. A spatial coordinate system can be constructed based on these three standard calibration points, where the standard calibration point located at the center of the L-shaped calibration plate is the origin of the coordinate system. In this way, the accurate position of the coordinate system can be determined, and the accurate relative position of the object to be measured from the coordinate system can be obtained.

[0087] The calibration board, calibrated using the above method, can operate at the same frequency as the high-speed camera, completing calibration image acquisition in a very short time. It is also self-illuminating, making it suitable for high-speed, short-exposure experimental scenarios. As a non-planar structure, the calibration board does not require repositioning during calibration; the mapping relationship between the spatial coordinates of the calibration points and the image coordinates can be obtained based on the positional relationships of different calibration points, thus completing the calibration. At least three standard calibration points on the board can be used to automatically identify and generate a coordinate system. After calibration using this method, the resulting spatial coordinate system is located at the same position as the calibration board, facilitating subsequent data processing.

[0088] In the above technical solution, this method fully utilizes the self-illumination, brightness encoding, and reference point characteristics of the device to achieve full automation from image acquisition, feature recognition, coordinate matching to parameter solving and coordinate system establishment. The entire method requires no movement of the calibration board and can be completed quickly under short exposure conditions, making it particularly suitable for the rapid deployment and on-site calibration of high-speed vision measurement systems, significantly improving calibration efficiency and ease of use. Simultaneously, by binding the observed set of luminous points with the known prior power of the calibration board in each frame, two types of consistency are constructed. Through frame-by-frame reliability judgment and robust solution mechanisms, the specific problem of difficulty in distinguishing calibration points caused by noise in high-speed short exposures is solved.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A stereo calibration device adapted to high-speed binocular vision scenarios, characterized in that: The device includes The L-shaped calibration plate body is composed of two calibration plates that intersect perpendicularly, and the calibration surfaces of the two calibration plates are provided with a preset calibration pattern composed of multiple self-illuminating calibration points. Among the multiple self-luminous calibration points, there is a standard calibration point that serves as a brightness reference, and there is a known and distinct difference between the luminous power of the remaining self-luminous calibration points and the luminous power of the standard calibration point; The control module is electrically connected to the self-illuminating calibration point and is used to control the emission timing of the self-illuminating calibration point so that its flashing frequency is synchronized with the shooting frame rate of the high-speed binocular camera.

2. The stereo calibration device according to claim 1, characterized in that: The self-illuminating calibration point is an LED bead or a light-emitting component made of luminescent material.

3. The stereo calibration device according to claim 2, characterized in that: The plurality of self-illuminating calibration points include at least three standard calibration points for constructing a spatial three-dimensional coordinate system in the calibration image; including the center point located at the connection of two calibration plates, the edge point of the connection, and the center point of the edge of a calibration plate away from the connection.

4. The stereo calibration device according to claim 3, characterized in that: The control module dynamically designates selected self-illuminating calibration points as standard calibration points for constructing the spatial three-dimensional coordinate system by setting a luminous intensity higher than a predetermined brightness threshold.

5. The stereo calibration device according to claim 1, characterized in that: The preset pattern is a rectangular dot matrix arranged at equal intervals.

6. A stereo calibration method for high-speed binocular vision scenes based on the device described in any one of claims 1-5, characterized in that, Includes the following steps: Step S1: Synchronously start the high-speed binocular camera and the stereo calibration device. The high-speed binocular camera acquires multiple frames of calibration images by combining different self-illuminating calibration points that are controlled by the control module to light up in a time-division manner. The control module controls the flashing frequency of the self-illuminating calibration points to be synchronized with the frame rate of the high-speed binocular camera. Step S2: Determine the position of the standard calibration point in the calibration image based on the gray value of the light spot, and determine the candidate region of each light emission calibration point based on the position of the standard calibration point and the arrangement structure of the self-emitting calibration points; Step S3: Evaluate the noise intensity of the current frame calibration image based on the distinguishability of the candidate region of each self-illuminating calibration point in the current frame calibration image and the observational consistency of the luminous power of all self-illuminating calibration points in the current frame calibration image; Step S4: Based on the noise intensity, select a reliable calibration subset from the multi-frame calibration images, and solve the calibration parameters of the high-speed binocular vision system based on the reliable calibration subset.

7. The stereo calibration method according to claim 6, characterized in that: Step S3, which assesses the noise intensity of the current frame calibration image based on the distinguishability of the candidate region of each self-illuminating calibration point in the current frame calibration image and the observational consistency of the luminous power of all self-illuminating calibration points in the current frame calibration image, includes: Step S31: Calculate the intramodal consistency index of the current frame calibration image based on the grayscale information and spot morphology information of the self-illuminating calibration points and their neighboring self-illuminating calibration points in the current frame calibration image. The intramodal consistency index is used to characterize the degree of distinguishability between the self-illuminating calibration points and their neighboring self-illuminating calibration points in the current frame calibration image. Step S32: Construct an observed grayscale vector based on the grayscale values ​​of all self-illuminating calibration points in the current frame calibration image, and construct a prior power vector based on the difference in luminous power of all self-illuminating calibration points; calculate the cosine similarity between the observed grayscale vector and the prior power vector to obtain an intermodal consistency index, which is used to characterize the degree of observational consistency between the grayscale distribution of the current frame image and the luminous power distribution of the calibration board; Step S33: Construct noise intensity by weighted fusion of the intramodal consistency index and the intermodal consistency index.

8. The stereo calibration method according to claim 7, characterized in that: Step S31, which calculates the intramodal consistency index of the current frame calibration image based on the grayscale information and spot morphology information of the self-illuminating calibration points and their neighboring self-illuminating calibration points in the current frame calibration image, includes: Step S311: Select a local background region outside the candidate region of each light emission calibration point, calculate the background mean based on the pixel gray value in the local background region, and estimate the local background noise variance based on the fluctuation of each pixel in the local background region relative to the background mean. Step S312: Normalize the peak gray value of the candidate region of the self-illuminating calibration point after background removal based on the background mean by using the local background noise variance to obtain the single-point signal-to-noise ratio of the self-illuminating calibration point. Step S313: Based on the minimum difference between the peak gray level of the candidate region of the self-illuminating calibration point and the peak gray level of the candidate region of its neighboring self-illuminating calibration points, the normalized value under the local background noise variance is obtained to obtain the neighboring point separation degree of the self-illuminating calibration point. Step S314: Collect all reference frames when the self-propelled cursor is fixed and lit, and obtain the spot shape anomaly degree based on the difference between the spot width of the self-propelled cursor fixed point in the current frame and the average spot width of the reference frame. Step S315: Aggregate the single-point signal-to-noise ratio, the adjacent point separation degree, and the spot shape anomaly degree to obtain the intramodal consistency index of the current frame calibration image.

9. The stereo calibration method according to claim 8, characterized in that: Step S32, which involves constructing an observed grayscale vector based on the grayscale values ​​of all self-illuminating calibration points in the current frame calibration image and constructing a priori power vector based on the differences in luminous power among all self-illuminating calibration points, includes: Step S321: In the current frame calibration image, remove the background from the peak gray value of each self-propelled cursor point and normalize it through the standard calibration point. Then, concatenate the peak gray values ​​of all the self-propelled cursor points after normalization as the observed gray value vector. Step S322: Normalize the luminous power of each self-firing cursor point to the luminous power of the standard calibration point, and concatenate the luminous power of all the normalized self-firing cursor points in sequence as a priori power vector.

10. The stereo calibration method according to claim 9, characterized in that: Step S34, which involves filtering and obtaining a reliable calibration subset from multiple calibration images based on the noise intensity, includes: The noise intensity of the calibration image is statistically modeled frame by frame using a Gaussian mixture model, which includes at least two Gaussian distribution components. One component represents the noise intensity distribution of reliable frames, and the other component represents the noise intensity distribution of noisy frames. The posterior probability of each calibration image belonging to a reliable calibration subset is obtained by calculating the probability density of the noise intensity of each frame under each component.

Images